b10545 b10541
ggml-org/llama.cppAug 21, 2026 → Aug 21, 20260 days between releases
What's New
Release order is inverted: b10541 predates b10545, so this 'upgrade' loses the Metal K-clamp fix.
New Features
- Adds a --mmproj-device CLI flag (short form -mmdev) for choosing which backend device hosts the multimodal projection model.
- The mmproj device backend is now loaded immediately at startup rather than lazily.
Improvements
- The MTMD_BACKEND_DEVICE environment variable is still honored, so existing setups keep working alongside the new flag.
- CLI documentation was regenerated to cover the new argument.
Upgrade Notes
b10541 has a lower build number and an earlier publish timestamp than b10545, so moving to it is effectively a downgrade and it does not contain the Metal Tensor API K-extent clamp fix. If you run Metal mat-mat with a K dimension that is not a multiple of 32, stay on b10545 or later to avoid the out-of-bounds reads and NaN results. The new --mmproj-device flag is purely additive and the MTMD_BACKEND_DEVICE env var remains supported, so no config migration is required.
The Two Releases
b10545
Aug 21, 2026· 2 months ago
Fixes a critical bug in the Metal Tensor API mat-mat kernel where reading out-of-bounds memory caused NaNs when the K dimension was not a multiple of 32.
Full release notesb10541
Aug 21, 2026· 2 months ago
Adds a command-line argument `--mmproj-device` to the MTMD (Multi-Modal Text-to-Model) backend to allow specifying the device for the multimodal projection model.
Full release notes